Avi Mendelson

Tal Rozen, Moshe Kimhi, Brian Chmiel et al.

Binary Neural Networks (BNNs) are an extremely promising method to reduce deep neural networks' complexity and power consumption massively. Binarization techniques, however, suffer from ineligible performance degradation compared to their full-precision counterparts. Prior work mainly focused on strategies for sign function approximation during forward and backward phases to reduce the quantization error during the binarization process. In this work, we propose a Bi-Modal Distributed binarization method (\methodname{}). That imposes bi-modal distribution of the network weights by kurtosis regularization. The proposed method consists of a training scheme that we call Weight Distribution Mimicking (WDM), which efficiently imitates the full-precision network weight distribution to their binary counterpart. Preserving this distribution during binarization-aware training creates robust and informative binary feature maps and significantly reduces the generalization error of the BNN. Extensive evaluations on CIFAR-10 and ImageNet demonstrate the superiority of our method over current state-of-the-art schemes. Our source code, experimental settings, training logs, and binary models are available at \url{https://github.com/BlueAnon/BD-BNN}.

Moshe Kimhi, Tal Rozen, Avi Mendelson et al.

Quantized neural networks are well known for reducing the latency, power consumption, and model size without significant harm to the performance. This makes them highly appropriate for systems with limited resources and low power capacity. Mixed-precision quantization offers better utilization of customized hardware that supports arithmetic operations at different bitwidths. Quantization methods either aim to minimize the compression loss given a desired reduction or optimize a dependent variable for a specified property of the model (such as FLOPs or model size); both make the performance inefficient when deployed on specific hardware, but more importantly, quantization methods assume that the loss manifold holds a global minimum for a quantized model that copes with the global minimum of the full precision counterpart. Challenging this assumption, we argue that the optimal minimum changes as the precision changes, and thus, it is better to look at quantization as a random process, placing the foundation for a different approach to quantize neural networks, which, during the training procedure, quantizes the model to a different precision, looks at the bit allocation as a Markov Decision Process, and then, finds an optimal bitwidth allocation for measuring specified behaviors on a specific device via direct signals from the particular hardware architecture. By doing so, we avoid the basic assumption that the loss behaves the same way for a quantized model. Automatic Mixed-Precision Quantization for Edge Devices (dubbed AMED) demonstrates its superiority over current state-of-the-art schemes in terms of the trade-off between neural network accuracy and hardware efficiency, backed by a comprehensive evaluation.

Amit LeVi, Raz Lapid, Rom Himelstein et al.

Large Language Models (LLMs) excel across diverse tasks, yet many applications require only limited capabilities, making large variants inefficient in memory and latency. Existing approaches often combine distillation and quantization, but most post-training quantization (PTQ) methods are task-agnostic, ignoring how task-specific signals are distributed across layers. In this work, we propose to use hidden representations that encode task-salient signals as a guideline for quantization. In order to fully utilize our innovative idea, this paper compares two new task-aware PTQ methods: Task-Aware Quantization (TAQ), which allocates bitwidths using task-conditioned statistics from hidden activations, and TAQO, which allocates precision based on direct layer sensitivity tests. From a small calibration set, these approaches identify task-relevant layers, preserving their precision while aggressively quantizing the rest. This yields stable task sensitivity profiles and efficient task-specialized models. Across models, TAQ and TAQO outperform the baselines; TAQ leads on Phi-4, while TAQO leads on Llama-3.1, Qwen3, and Qwen2.5. For instances, on Phi-4 it achieves 42.33 EM / 50.81 F1, far surpassing Activation-aware Weight Quantization (AWQ) (2.25 / 7.07), while remaining within < 1.0% of the original accuracy at lower average precision.

Rom Himelstein, Amit LeVi, Brit Youngmann et al.

Safety-aligned large language models (LLMs) are becoming increasingly widespread, especially in sensitive applications where fairness is essential and biased outputs can cause significant harm. However, evaluating the fairness of models is a complex challenge, and approaches that do so typically utilize standard question-answer (QA) styled schemes. Such methods often overlook deeper issues by interpreting the model's refusal responses as positive fairness measurements, which creates a false sense of fairness. In this work, we introduce the concept of silenced biases, which are unfair preferences encoded within models' latent space and are effectively concealed by safety-alignment. Previous approaches that considered similar indirect biases often relied on prompt manipulation or handcrafted implicit queries, which present limited scalability and risk contaminating the evaluation process with additional biases. We propose the Silenced Bias Benchmark (SBB), which aims to uncover these biases by employing activation steering to reduce model refusals during QA. SBB supports easy expansion to new demographic groups and subjects, presenting a fairness evaluation framework that encourages the future development of fair models and tools beyond the masking effects of alignment training. We demonstrate our approach over multiple LLMs, where our findings expose an alarming distinction between models' direct responses and their underlying fairness issues.

Amit Levi, Rom Himelstein, Yaniv Nemcovsky et al.

Safety-aligned LLMs respond to prompts with either compliance or refusal, each corresponding to distinct directions in the model's activation space. Recent works show that initializing attacks via self-transfer from other prompts significantly enhances their performance. However, the underlying mechanisms of these initializations remain unclear, and attacks utilize arbitrary or hand-picked initializations. This work presents that each gradient-based jailbreak attack and subsequent initialization gradually converge to a single compliance direction that suppresses refusal, thereby enabling an efficient transition from refusal to compliance. Based on this insight, we propose CRI, an initialization framework that aims to project unseen prompts further along compliance directions. We demonstrate our approach on multiple attacks, models, and datasets, achieving an increased attack success rate (ASR) and reduced computational overhead, highlighting the fragility of safety-aligned LLMs. A reference implementation is available at: https://amit1221levi.github.io/CRI-Jailbreak-Init-LLMs-evaluation.

Elias Shaheen, Avi Mendelson

The general-purpose nature of Large Language Models (LLMs) presents a significant challenge for domain-specific applications, often leading to out-of-domain (OOD) interactions that undermine the provider's intent. Existing methods for detecting such scenarios treat the LLM as an uninterpretable black box and overlook the internal processing of inputs. In this work we show that layer transitions provide a promising avenue for extracting domain-specific signature. Specifically, we present several lightweight ways of learning on internal dynamics encoded using a sparse autoencoder (SAE) that exhibit great capability in distinguishing OOD texts. Building on top of SAEs representation transitions enables us to better interpret the LLM internal evolution of input processing and shed light on its decisions. We provide a comprehensive analysis of the method and benchmark it with the gemma-2 2B and 9B models. Our results emphasize the efficacy of the internal process in capturing fine-grained input-related details.

Rom Himelstein, Amit LeVi, Yonatan Belinkov et al.

Padding tokens are widely used in large language models (LLMs) to equalize sequence lengths during batched inference. While they should be fully masked, implementation errors can cause them to influence computation, and the extent of this influence is not well understood. We systematically study this effect across three open-source model families (Llama, Gemma, Qwen), inserting controlled amounts of padding and evaluating outcomes along four axes: activations, generation quality, bias, and safety. Even small amounts of padding shift hidden representations, degrade quality in smaller models, alter bias in unpredictable ways, and weaken safety guardrails. These findings demonstrate that padding is not a harmless detail but a robustness risk that must be carefully handled in deployment.

Yaniv Nemcovsky, Avi Mendelson, Chaim Baskin

Sparse and patch adversarial attacks were previously shown to be applicable in realistic settings and are considered a security risk to autonomous systems. Sparse adversarial perturbations constitute a setting in which the adversarial perturbations are limited to affecting a relatively small number of points in the input. Patch adversarial attacks denote the setting where the sparse attacks are limited to a given structure, i.e., sparse patches with a given shape and number. However, previous patch adversarial attacks do not simultaneously optimize multiple patches' locations and perturbations. This work suggests a novel approach for sparse patches adversarial attacks via point-wise trimming dense adversarial perturbations. Our approach enables simultaneous optimization of multiple sparse patches' locations and perturbations for any given number and shape. Moreover, our approach is also applicable for standard sparse adversarial attacks, where we show that it significantly improves the state-of-the-art over multiple extensive settings. A reference implementation of the proposed method and the reported experiments is provided at \url{https://github.com/yanemcovsky/SparsePatches.git}

Or Feldman, Amit Boyarski, Shai Feldman et al.

Graph isomorphism testing is usually approached via the comparison of graph invariants. Two popular alternatives that offer a good trade-off between expressive power and computational efficiency are combinatorial (i.e., obtained via the Weisfeiler-Leman (WL) test) and spectral invariants. While the exact power of the latter is still an open question, the former is regularly criticized for its limited power, when a standard configuration of uniform pre-coloring is used. This drawback hinders the applicability of Message Passing Graph Neural Networks (MPGNNs), whose expressive power is upper bounded by the WL test. Relaxing the assumption of uniform pre-coloring, we show that one can increase the expressive power of the WL test ad infinitum. Following that, we propose an efficient pre-coloring based on spectral features that provably increase the expressive power of the vanilla WL test. The above claims are accompanied by extensive synthetic and real data experiments. The code to reproduce our experiments is available at https://github.com/TPFI22/Spectral-and-Combinatorial

Maxim Fishman, Chaim Baskin, Evgenii Zheltonozhskii et al.

Graph neural networks (GNNs) have become a powerful tool for processing graph-structured data but still face challenges in effectively aggregating and propagating information between layers, which limits their performance. We tackle this problem with the kernel regression (KR) approach, using KR loss as the primary loss in self-supervised settings or as a regularization term in supervised settings. We show substantial performance improvements compared to state-of-the-art in both scenarios on multiple transductive and inductive node classification datasets, especially for deep networks. As opposed to mutual information (MI), KR loss is convex and easy to estimate in high-dimensional cases, even though it indirectly maximizes the MI between its inputs. Our work highlights the potential of KR to advance the field of graph representation learning and enhance the performance of GNNs. The code to reproduce our experiments is available at https://github.com/Anonymous1252022/KR_for_GNNs

Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson et al.

The success of learning with noisy labels (LNL) methods relies heavily on the success of a warm-up stage where standard supervised training is performed using the full (noisy) training set. In this paper, we identify a "warm-up obstacle": the inability of standard warm-up stages to train high quality feature extractors and avert memorization of noisy labels. We propose "Contrast to Divide" (C2D), a simple framework that solves this problem by pre-training the feature extractor in a self-supervised fashion. Using self-supervised pre-training boosts the performance of existing LNL approaches by drastically reducing the warm-up stage's susceptibility to noise level, shortening its duration, and improving extracted feature quality. C2D works out of the box with existing methods and demonstrates markedly improved performance, especially in the high noise regime, where we get a boost of more than 27% for CIFAR-100 with 90% noise over the previous state of the art. In real-life noise settings, C2D trained on mini-WebVision outperforms previous works both in WebVision and ImageNet validation sets by 3% top-1 accuracy. We perform an in-depth analysis of the framework, including investigating the performance of different pre-training approaches and estimating the effective upper bound of the LNL performance with semi-supervised learning. Code for reproducing our experiments is available at https://github.com/ContrastToDivide/C2D

Evgenii Zheltonozhskii, Chaim Baskin, Alex M. Bronstein et al.

Unsupervised learning has always been appealing to machine learning researchers and practitioners, allowing them to avoid an expensive and complicated process of labeling the data. However, unsupervised learning of complex data is challenging, and even the best approaches show much weaker performance than their supervised counterparts. Self-supervised deep learning has become a strong instrument for representation learning in computer vision. However, those methods have not been evaluated in a fully unsupervised setting. In this paper, we propose a simple scheme for unsupervised classification based on self-supervised representations. We evaluate the proposed approach with several recent self-supervised methods showing that it achieves competitive results for ImageNet classification (39% accuracy on ImageNet with 1000 clusters and 46% with overclustering). We suggest adding the unsupervised evaluation to a set of standard benchmarks for self-supervised learning. The code is available at https://github.com/Randl/kmeans_selfsuper

Yaniv Nemcovsky, Evgenii Zheltonozhskii, Chaim Baskin et al.

Deep neural networks are known to be vulnerable to adversarial attacks. Current methods of defense from such attacks are based on either implicit or explicit regularization, e.g., adversarial training. Randomized smoothing, the averaging of the classifier outputs over a random distribution centered in the sample, has been shown to guarantee the performance of a classifier subject to bounded perturbations of the input. In this work, we study the application of randomized smoothing as a way to improve performance on unperturbed data as well as to increase robustness to adversarial attacks. The proposed technique can be applied on top of any existing adversarial defense, but works particularly well with the randomized approaches. We examine its performance on common white-box (PGD) and black-box (transfer and NAttack) attacks on CIFAR-10 and CIFAR-100, substantially outperforming previous art for most scenarios and comparable on others. For example, we achieve 60.4% accuracy under a PGD attack on CIFAR-10 using ResNet-20, outperforming previous art by 11.7%. Since our method is based on sampling, it lends itself well for trading-off between the model inference complexity and its performance. A reference implementation of the proposed techniques is provided at https://github.com/yanemcovsky/SIAM

Yury Nahshan, Brian Chmiel, Chaim Baskin et al.

Neural network quantization enables the deployment of large models on resource-constrained devices. Current post-training quantization methods fall short in terms of accuracy for INT4 (or lower) but provide reasonable accuracy for INT8 (or above). In this work, we study the effect of quantization on the structure of the loss landscape. Additionally, we show that the structure is flat and separable for mild quantization, enabling straightforward post-training quantization methods to achieve good results. We show that with more aggressive quantization, the loss landscape becomes highly non-separable with steep curvature, making the selection of quantization parameters more challenging. Armed with this understanding, we design a method that quantizes the layer parameters jointly, enabling significant accuracy improvement over current post-training quantization methods. Reference implementation is available at https://github.com/ynahshan/nn-quantization-pytorch/tree/master/lapq

Chaim Baskin, Brian Chmiel, Evgenii Zheltonozhskii et al.

Convolutional neural networks (CNNs) have become the dominant neural network architecture for solving visual processing tasks. One of the major obstacles hindering the ubiquitous use of CNNs for inference is their relatively high memory bandwidth requirements, which can be a main energy consumer and throughput bottleneck in hardware accelerators. Accordingly, an efficient feature map compression method can result in substantial performance gains. Inspired by quantization-aware training approaches, we propose a compression-aware training (CAT) method that involves training the model in a way that allows better compression of feature maps during inference. Our method trains the model to achieve low-entropy feature maps, which enables efficient compression at inference time using classical transform coding methods. CAT significantly improves the state-of-the-art results reported for quantization. For example, on ResNet-34 we achieve 73.1% accuracy (0.2% degradation from the baseline) with an average representation of only 1.79 bits per value. Reference implementation accompanies the paper at https://github.com/CAT-teams/CAT

Brian Chmiel, Chaim Baskin, Ron Banner et al.

Convolutional neural networks (CNNs) achieve state-of-the-art accuracy in a variety of tasks in computer vision and beyond. One of the major obstacles hindering the ubiquitous use of CNNs for inference on low-power edge devices is their high computational complexity and memory bandwidth requirements. The latter often dominates the energy footprint on modern hardware. In this paper, we introduce a lossy transform coding approach, inspired by image and video compression, designed to reduce the memory bandwidth due to the storage of intermediate activation calculation results. Our method does not require fine-tuning the network weights and halves the data transfer volumes to the main memory by compressing feature maps, which are highly correlated, with variable length coding. Our method outperform previous approach in term of the number of bits per value with minor accuracy degradation on ResNet-34 and MobileNetV2. We analyze the performance of our approach on a variety of CNN architectures and demonstrate that FPGA implementation of ResNet-18 with our approach results in a reduction of around 40% in the memory energy footprint, compared to quantized network, with negligible impact on accuracy. When allowing accuracy degradation of up to 2%, the reduction of 60% is achieved. A reference implementation is available at https://github.com/CompressTeam/TransformCodingInference

Yochai Zur, Chaim Baskin, Evgenii Zheltonozhskii et al.

Recently, deep learning has become a de facto standard in machine learning with convolutional neural networks (CNNs) demonstrating spectacular success on a wide variety of tasks. However, CNNs are typically very demanding computationally at inference time. One of the ways to alleviate this burden on certain hardware platforms is quantization relying on the use of low-precision arithmetic representation for the weights and the activations. Another popular method is the pruning of the number of filters in each layer. While mainstream deep learning methods train the neural networks weights while keeping the network architecture fixed, the emerging neural architecture search (NAS) techniques make the latter also amenable to training. In this paper, we formulate optimal arithmetic bit length allocation and neural network pruning as a NAS problem, searching for the configurations satisfying a computational complexity budget while maximizing the accuracy. We use a differentiable search method based on the continuous relaxation of the search space proposed by Liu et al. (arXiv:1806.09055). We show, by grid search, that heterogeneous quantized networks suffer from a high variance which renders the benefit of the search questionable. For pruning, improvement over homogeneous cases is possible, but it is still challenging to find those configurations with the proposed method. The code is publicly available at https://github.com/yochaiz/Slimmable and https://github.com/yochaiz/darts-UNIQ

Chaim Baskin, Natan Liss, Yoav Chai et al.

Convolutional Neural Networks (CNN) are very popular in many fields including computer vision, speech recognition, natural language processing, to name a few. Though deep learning leads to groundbreaking performance in these domains, the networks used are very demanding computationally and are far from real-time even on a GPU, which is not power efficient and therefore does not suit low power systems such as mobile devices. To overcome this challenge, some solutions have been proposed for quantizing the weights and activations of these networks, which accelerate the runtime significantly. Yet, this acceleration comes at the cost of a larger error. The \uniqname method proposed in this work trains quantized neural networks by noise injection and a learned clamping, which improve the accuracy. This leads to state-of-the-art results on various regression and classification tasks, e.g., ImageNet classification with architectures such as ResNet-18/34/50 with low as 3-bit weights and activations. We implement the proposed solution on an FPGA to demonstrate its applicability for low power real-time applications. The implementation of the paper is available at https://github.com/Lancer555/NICE

Liran Cohen, Yaniv Nemcovesky, Avi Mendelson

Machine unlearning aims to remove the influence of specific training data from a model without requiring full retraining. This capability is crucial for ensuring privacy, safety, and regulatory compliance. Therefore, verifying whether a model has truly forgotten target data is essential for maintaining reliability and trustworthiness. However, existing evaluation methods often assess forgetting at the level of individual inputs. This approach may overlook residual influence present in semantically similar examples. Such influence can compromise privacy and lead to indirect information leakage. We propose REMIND (Residual Memorization In Neighborhood Dynamics), a novel evaluation method aiming to detect the subtle remaining influence of unlearned data and classify whether the data has been effectively forgotten. REMIND analyzes the model's loss over small input variations and reveals patterns unnoticed by single-point evaluations. We show that unlearned data yield flatter, less steep loss landscapes, while retained or unrelated data exhibit sharper, more volatile patterns. REMIND requires only query-based access, outperforms existing methods under similar constraints, and demonstrates robustness across different models, datasets, and paraphrased inputs, making it practical for real-world deployment. By providing a more sensitive and interpretable measure of unlearning effectiveness, REMIND provides a reliable framework to assess unlearning in language models. As a result, REMIND offers a novel perspective on memorization and unlearning.

Moshe Kimhi, Idan Kashani, Avi Mendelson et al.

The widely used ReLU is favored for its hardware efficiency, {as the implementation at inference is a one bit sign case,} yet suffers from issues such as the ``dying ReLU'' problem, where during training, neurons fail to activate and constantly remain at zero, as highlighted by Lu et al. Traditional approaches to mitigate this issue often introduce more complex and less hardware-friendly activation functions. In this work, we propose a Hysteresis Rectified Linear Unit (HeLU), an efficient activation function designed to address the ``dying ReLU'' problem with minimal complexity. Unlike traditional activation functions with fixed thresholds for training and inference, HeLU employs a variable threshold that refines the backpropagation. This refined mechanism allows simpler activation functions to achieve competitive performance comparable to their more complex counterparts without introducing unnecessary complexity or requiring inductive biases. Empirical evaluations demonstrate that HeLU enhances model generalization across diverse datasets, offering a promising solution for efficient and effective inference suitable for a wide range of neural network architectures.

Nate Rothschild, Moshe Kimhi, Avi Mendelson et al.

Image reconstruction from corrupted images is crucial across many domains. Most reconstruction networks are trained on post-ISP sRGB images, even though the image-signal-processing pipeline irreversibly mixes colors, clips dynamic range, and blurs fine detail. This paper uses the rain degradation problem as a use case to show that these losses are avoidable, and demonstrates that learning directly on raw Bayer mosaics yields superior reconstructions. To substantiate the claim, we (i) evaluate post-ISP and Bayer reconstruction pipelines, (ii) curate Raw-Rain, the first public benchmark of real rainy scenes captured in both 12-bit Bayer and bit-depth-matched sRGB, and (iii) introduce Information Conservation Score (ICS), a color-invariant metric that aligns more closely with human opinion than PSNR or SSIM. On the test split, our raw-domain model improves sRGB results by up to +0.99 dB PSNR and +1.2% ICS, while running faster with half of the GFLOPs. The results advocate an ISP-last paradigm for low-level vision and open the door to end-to-end learnable camera pipelines.

Idan Kashani, Avi Mendelson, Yaniv Nemcovsky

Large language models (LLMs) achieve impressive results over various tasks, and ever-expanding public repositories contain an abundance of pre-trained models. Therefore, identifying the best-performing LLM for a given task is a significant challenge. Previous works have suggested learning LLM representations to address this. However, these approaches present limited scalability and require costly retraining to encompass additional models and datasets. Moreover, the produced representation utilizes distinct spaces that cannot be easily interpreted. This work presents an efficient, training-free approach to representing LLMs as linear operators within the prompts' semantic task space, thus providing a highly interpretable representation of the models' application. Our method utilizes closed-form computation of geometrical properties and ensures exceptional scalability and real-time adaptability to dynamically expanding repositories. We demonstrate our approach on success prediction and model selection tasks, achieving competitive or state-of-the-art results with notable performance in out-of-sample scenarios.

Alex Karbachevsky, Chaim Baskin, Evgenii Zheltonozhskii et al.

Convolutional Neural Networks (CNNs) have become common in many fields including computer vision, speech recognition, and natural language processing. Although CNN hardware accelerators are already included as part of many SoC architectures, the task of achieving high accuracy on resource-restricted devices is still considered challenging, mainly due to the vast number of design parameters that need to be balanced to achieve an efficient solution. Quantization techniques, when applied to the network parameters, lead to a reduction of power and area and may also change the ratio between communication and computation. As a result, some algorithmic solutions may suffer from lack of memory bandwidth or computational resources and fail to achieve the expected performance due to hardware constraints. Thus, the system designer and the micro-architect need to understand at early development stages the impact of their high-level decisions (e.g., the architecture of the CNN and the amount of bits used to represent its parameters) on the final product (e.g., the expected power saving, area, and accuracy). Unfortunately, existing tools fall short of supporting such decisions. This paper introduces a hardware-aware complexity metric that aims to assist the system designer of the neural network architectures, through the entire project lifetime (especially at its early stages) by predicting the impact of architectural and micro-architectural decisions on the final product. We demonstrate how the proposed metric can help evaluate different design alternatives of neural network models on resource-restricted devices such as real-time embedded systems, and to avoid making design mistakes at early stages.

Evgenii Zheltonozhskii, Chaim Baskin, Yaniv Nemcovsky et al.

Even though deep learning has shown unmatched performance on various tasks, neural networks have been shown to be vulnerable to small adversarial perturbations of the input that lead to significant performance degradation. In this work we extend the idea of adding white Gaussian noise to the network weights and activations during adversarial training (PNI) to the injection of colored noise for defense against common white-box and black-box attacks. We show that our approach outperforms PNI and various previous approaches in terms of adversarial accuracy on CIFAR-10 and CIFAR-100 datasets. In addition, we provide an extensive ablation study of the proposed method justifying the chosen configurations.

Natan Liss, Chaim Baskin, Avi Mendelson et al.

Convolutional Neural Networks (CNN) has become more popular choice for various tasks such as computer vision, speech recognition and natural language processing. Thanks to their large computational capability and throughput, GPUs ,which are not power efficient and therefore does not suit low power systems such as mobile devices, are the most common platform for both training and inferencing tasks. Recent studies has shown that FPGAs can provide a good alternative to GPUs as a CNN accelerator, due to their re-configurable nature, low power and small latency. In order for FPGA-based accelerators outperform GPUs in inference task, both the parameters of the network and the activations must be quantized. While most works use uniform quantizers for both parameters and activations, it is not always the optimal one, and a non-uniform quantizer need to be considered. In this work we introduce a custom hardware-friendly approach to implement non-uniform quantizers. In addition, we use a single scale integer representation of both parameters and activations, for both training and inference. The combined method yields a hardware efficient non-uniform quantizer, fit for real-time applications. We have tested our method on CIFAR-10 and CIFAR-100 image classification datasets with ResNet-18 and VGG-like architectures, and saw little degradation in accuracy.

Chaim Baskin, Eli Schwartz, Evgenii Zheltonozhskii et al.

We present a novel method for neural network quantization that emulates a non-uniform $k$-quantile quantizer, which adapts to the distribution of the quantized parameters. Our approach provides a novel alternative to the existing uniform quantization techniques for neural networks. We suggest to compare the results as a function of the bit-operations (BOPS) performed, assuming a look-up table availability for the non-uniform case. In this setup, we show the advantages of our strategy in the low computational budget regime. While the proposed solution is harder to implement in hardware, we believe it sets a basis for new alternatives to neural networks quantization.

Chaim Baskin, Natan Liss, Evgenii Zheltonozhskii et al.

Deep neural networks (DNNs) are used by different applications that are executed on a range of computer architectures, from IoT devices to supercomputers. The footprint of these networks is huge as well as their computational and communication needs. In order to ease the pressure on resources, research indicates that in many cases a low precision representation (1-2 bit per parameter) of weights and other parameters can achieve similar accuracy while requiring less resources. Using quantized values enables the use of FPGAs to run NNs, since FPGAs are well fitted to these primitives; e.g., FPGAs provide efficient support for bitwise operations and can work with arbitrary-precision representation of numbers. This paper presents a new streaming architecture for running QNNs on FPGAs. The proposed architecture scales out better than alternatives, allowing us to take advantage of systems with multiple FPGAs. We also included support for skip connections, that are used in state-of-the art NNs, and shown that our architecture allows to add those connections almost for free. All this allowed us to implement an 18-layer ResNet for 224x224 images classification, achieving 57.5% top-1 accuracy. In addition, we implemented a full-sized quantized AlexNet. In contrast to previous works, we use 2-bit activations instead of 1-bit ones, which improves AlexNet's top-1 accuracy from 41.8% to 51.03% for the ImageNet classification. Both AlexNet and ResNet can handle 1000-class real-time classification on an FPGA. Our implementation of ResNet-18 consumes 5x less power and is 4x slower for ImageNet, when compared to the same NN on the latest Nvidia GPUs. Smaller NNs, that fit a single FPGA, are running faster then on GPUs on small (32x32) inputs, while consuming up to 20x less energy and power.